/kaggle-human-protein-atlas

Sharing code for Kaggle Human Protein Atlas competition. Top 4% in Leaderboard.

Primary LanguageJupyter Notebook

kaggle-human-protein-atlas

Sharing code for Kaggle Human Protein Atlas competition

Top 4% in Leaderboard

https://www.kaggle.com/c/human-protein-atlas-image-classification

Data

Color filters for the fluorescent images

  • Green (Alexa Flour 488): The protein of interest
  • Blue (DAPI): Nucleus
  • Red: Microtubules
  • Yellow: Endoplasmic reticulum

Train sample size: 31072

Test sample size: 11702

Labels / Cell component / # images in train set

  1. Nucleoplasm 12885
  2. Nuclear membrane 1254
  3. Nucleoli 3621
  4. Nucleoli fibrillar center 1561
  5. Nuclear speckles 1858
  6. Nuclear bodies 2513
  7. Endoplasmic reticulum 1008
  8. Golgi apparatus 2822
  9. Peroxisomes 53
  10. Endosomes 45
  11. Lysosomes 28
  12. Intermediate filaments 1093
  13. Actin filaments 688
  14. Focal adhesion sites 537
  15. Microtubules 1066
  16. Microtubule ends 21
  17. Cytokinetic bridge 530
  18. Mitotic spindle 210
  19. Microtubule organizing center 902
  20. Centrosome 1482
  21. Lipid droplets 172
  22. Plasma membrane 3777
  23. Cell junctions 802
  24. Mitochondria 2965
  25. Aggresome 322
  26. Cytosol 8228
  27. Cytoplasmic bodies 828
  28. Rods & rings 11

Quickstart

Resnet FastAI

python3 -m code.resnet_fastai --imagesize=256 --loss=focal --gpuid=0

python3 -m code.resnet_fastai --model=stage-2-resnet50-224-drop0.5-ep5_15 --gpuid=0

  • -a, --arch : Neural network architecture
  • -b, --batchsize : batch size
  • -d, --encoderdepth : encoder depth of the network
  • -D, --dataset : Dataset
  • -e, --epochnum1 : epoch number for stage 1
  • -E, --epochnum2 : epoch number for stage 2
  • -f, --fold : K fold cross validation
  • -i, --gpuid : GPU device id
  • -l, --loss : loss function
  • -m, --model : trained model to load
  • -p, --dropout : dropout ratio
  • -r, --learningrate : learning rate
  • -s, --size : image size
  • -S, --sampler : sampler
  • -t, --thres : threshold
  • -v, --verbose : set verbosity 0-3, 0 to turn off output (not yet implemented)

Mask RCNN

python3 -m code.mask_rcnn train --dataset=data/official --subset=train --weights=coco

python3 -m code.mask_rcnn detect --dataset=data/official --subset=test --weights=logs/nucleus20181212T0318/mask_rcnn_nucleus_0025.h5